The impact of covid 19 on the minimum nights of stay for airbnb’s in Amsterdam
Team 1
Introduction
Introduction
For the project of the course Data Preparation and Workflow Management at Tilburg University, we decided to analyze the Airbnb market in Amsterdam and especially if the COVID-19 pandemic had an influence on the required minimum nights of stay. Especially since there are contradicted foundings in literature regarding this subject. A recent article by the New York Times suggested that the minimum nights of stay increased in New York City during the COVID-19 pandemic, whereas research by Kourtit et al. concluded that the minimum night requirements actually decreased during the pandemic. We decided to take a further look at these contradictions, by researching this subject. We collected data from Airbnb in Amsterdam, from 2020 as well as 2022, to see if there is any significant difference in the minimum nights of stay between during and after the COVID-19 pandemic.
Motivation
Motivation
Samengevoegd met Intro
Research Method
Research Method (1)
We decided to run a linear regression on the variables of interest. The dependent variable, the required minimum nights, is a metric variable and the independent variable, the presence of COVID-19 (present vs. absent) is a non-metric variable. We have data from 2020 and 2022 for 3960 different Airbnb listings (in total 7920 observations). The variable gets the value 1 assigned if the data is from 2020, so when there was COVID-19 in the Netherlands. Following from that, the variable gets the value 0 assigned if the data is from 2022, when the COVID-19 pandemic no longer had far-reaching consequences in the Netherlands. We decided to not only include the minimum nights of stay and the presence of COVID-19, but also added some control variables to our analysis, to see if there are other effects that might play a role. Since these control variables are differing in metric and non-metric variables, we have chosen linear regression over an ANOVA-analysis.
Research method (2)
Next to the dependent variable, the minimum_nights, and the independent variable covid, we included some control variables in a first regression. The control variables neighbourhood_num and roomtype_num were converted to factors, in which each number represents a different neighbourhood or roomtype. Next to that, accomodates, price and instant_bookable were included in this regression.
Before any conclusions can be drawn, we need to perform some robustness checks.
Independence (1)
# create a scatterplot of the residuals against the predicted values from the linear regression model 'm1'plot(m1$fitted.values, m1$residuals, xlab ="Fitted Values", ylab ="Residuals", main ="Residuals vs. Fitted Values Plot",ylim =c(-50, 60))abline(h =0, lty =2, col ='red')
Independence (1)
A first option to check for independence, is to create a scatterplot of the residuals against the fitted values from the linear regression model (in this case m1). The residuals should be independent from the variable, but this scatterplot shows us that this is not the case. We can conclude that there is no independence of the residuals.
Independence (2)
# Create a scatterplot of predicted vs actual valuesggplot(df_cleaned, aes(x = predicted, y = minimum_nights)) +geom_point() +# adds points to the plotgeom_abline(intercept =0, slope =1, color ="red") +# adds a diagonal line to the plot to visualize where predicted = actualxlab("Predicted Values") +# adds a label for the x-axisylab("Actual Values") +# adds a label for the y-axisggtitle("Predicted vs Actual Values Plot") # adds a title to the plot
Independence (2)
Next to that, we can create a scatterplot of the predicted values against the actual values.
Independence (3)
# perform Durbin-Watson test
Independence (3)
# perform Durbin-Watson testdwtest(m1)
Durbin-Watson test
data: m1
DW = 0.1109, p-value < 2.2e-16
alternative hypothesis: true autocorrelation is greater than 0
Independence (3)
# perform Durbin-Watson testdwtest(m1)
Durbin-Watson test
data: m1
DW = 0.1109, p-value < 2.2e-16
alternative hypothesis: true autocorrelation is greater than 0
TYPE HIER: a last thing we can do to check for independence is performing a Durbin-Watson test…
Homoskedasticity (1)
We can use the same plots when checking for homoskedasticity, as for checking independence.
Homoskedasticity (2)
# perform Breusch-Pagan test
Homoskedasticity (2)
# perform Breusch-Pagan testbptest(m1)
studentized Breusch-Pagan test
data: m1
BP = 93.6, df = 28, p-value = 5.38e-09
Homoskedasticity (2)
# perform Breusch-Pagan testbptest(m1)
studentized Breusch-Pagan test
data: m1
BP = 93.6, df = 28, p-value = 5.38e-09
TYPE HIER: a last thing we can do to check for homoskedasticity is performing a Breusch-Pagan test…
Normality (1)
## Making a dataframe with the residualsresiduals <-resid(m1)residuals_df <-data.frame(residuals = residuals)# Test for normality of residuals with a histogramggplot(residuals_df, aes(x = residuals)) +geom_histogram(binwidth =0.5, color ="black", fill ="white") +xlab("Residuals") +ylab("Frequency") +ggtitle("Histogram of Residuals")
Normality (1)
TYPE HIER: Type hier waarom de graph laat zien dat er geen independence is
Normality (2)
# Test for normality of residuals with a density plotggdensity(residuals_df$residuals, main ="Density plot of residuals",xlab ="residuals")
Normality (2)
TYPE HIER: Type hier waarom de graph laat zien dat er geen independence is
Normality (3)
# Test for normality with a Q-Q plotqqnorm(residuals)qqline(residuals)
Normality (3)
TYPE HIER: Type hier waarom de graph laat zien dat er geen independence is
Normality (4)
# Create random subsample of 5000 observations, so we are able to run a Shapiro-Wilk normality test (5000 is the maximum sample size)set.seed(123)my_subsample <- residuals_df[sample(nrow(residuals_df), 5000), ]shapiro.test(my_subsample)
Shapiro-Wilk normality test
data: my_subsample
W = 0.48635, p-value < 2.2e-16
TYPE HIER: a last thing we can do to check for normality is performing a Shapiro-Wilk normality test…
Linearity (1)
We can use the same first plot as used for testing independence and homoskedasticity. We can conclude…
